What the documents look like
In the complaints and official replies category, copy drafting for product and marketing statements draws on material from the 12315 platform, platform complaint centers, customer service chat, and order systems. The documents take the form of complaint tickets, regulatory inquiry letters, order evidence, and rectification reports. Their update frequency is irregular and case by case. Each case revolves around a specific transaction and covers product page promotional sentences, campaign rules, order status, and reply wording.
When drafting copy, the relevant original passages must be retrieved from this material and checked against the statement being prepared. Retrieval results directly shape the textual basis of a rectification reply. Where material is missing, the gap must be recorded. The regulatory-facing nature of this category means the retrieved passages carry formal evidentiary weight, so the drafting step depends on locating the exact original wording rather than paraphrasing a general impression.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be associated together. The vector model identifier must be verified first, and non-matching bases excluded and recorded as an observation. Cross-model retrieval is not executed in order to compare missed recalls.
Search mode is compared across semantic retrieval, full-text retrieval, and hybrid retrieval using the same test question. No single mode is presumed suitable for all material. The citation cap is measured in tokens. Material of different lengths is used to test whether content is returned completely, and the cap is then adjusted. The minimum relevance score takes effect only under semantic retrieval or when result reranking is used. Relevant and irrelevant samples are labelled before the threshold is tuned. Question optimization performs coreference resolution and question expansion. The original question and the optimized question are compared to check whether the subject has been altered.
Retrieval is a means of obtaining material. It does not perform permission or entity isolation. A similarity score does not exclude unrelated business records or unauthorized records. Hybrid retrieval merges multiple result paths through RRF, which is a fusion process rather than a switch. Image retrieval requires a corresponding image input and model capability, and reranking requires a model and must be enabled according to settings. The optional paths in the overall retrieval flow are not all executed on every request.
Hypothetical paper exercise sample: the source material reads, "This store ran a spend-and-save campaign from March 1 to March 15, 2025, and the campaign rules are subject to the notice published on the product page. Order A-20250310 was placed on March 10, and the amount actually paid was the campaign price." The draft sentence to be assessed reads, "This order has completed the discount according to the campaign rules, and there is no unfulfilled commitment." The check is whether "has completed the discount according to the campaign rules" is supported by the source material. The material states only the order date, the amount actually paid, and the campaign period. It does not state the fulfillment status of the discount or the content of any commitment.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Select the complaint ticket base and the order evidence base that carry the same vector model identifier | Verify the vector model identifier, exclude bases with different models, and do not execute cross-model retrieval |
| Search mode | Test semantic retrieval and full-text retrieval separately | Use the same test question to compare whether the passages returned by each mode cover the subject and the period |
| Citation cap | Hypothetical starting point of 2000 tokens | Use material of different lengths to test whether content is complete, and compare against a shorter and a longer value |
| Minimum relevance score | Hypothetical starting point of 0.5 | Label relevant and irrelevant samples first, then compare whether irrelevant samples are filtered at different thresholds |
| Question optimization | Enable and record the optimized question | Compare the original question with the optimized question and check whether the subject, object, and period have been altered |
Three common errors
- Associating knowledge bases that use different vector models. The input is two bases whose vector model identifiers differ. The comparison is the identifier field, and the observable result is whether the non-matching base is excluded. If the identifier is not verified before association, the retrieval scope includes a base that does not meet the condition.
- Treating the minimum relevance score as a permission or entity filter. The input is an unrelated business record. The comparison is the similarity score against the threshold, and the observable result is whether the unrelated record is filtered because its score is low. Relevance does not exclude unauthorized records, and access permission must be checked separately.
- Using the optimized retrieval question directly for the judgment. The input is the original question and the optimized question. The comparison is the subject and period in each, and the observable result is whether optimization introduces an object that the original question did not contain. The optimized question must be checked item by item against the source material.
How to verify the setup
- For the hypothetical exercise sample, input the source material and the draft sentence, then compare "has completed the discount according to the campaign rules" against the order date, the amount actually paid, and the campaign period stated in the material. Because the material does not state the fulfillment status of the discount or the content of any commitment, the draft sentence remains unsupported, and the evidence gap must be recorded.
- Input an explicit paper empty result of []. Check whether the downstream design follows the empty-reference path. Do not claim that a fictional rule retrieval necessarily returns []. Carrying an empty result requires an explicit paper input and a downstream handling design.
- For cross-model base association, input two bases whose vector model identifiers differ. Compare the identifier field and observe whether the association condition is met. Exclude the non-matching base and record the observation. Do not fabricate the runtime result of associating them together.
The primary task on this page is whether a specific promotional sentence in product copy is supported by the source material. It is not an enterprise identity certification tutorial, and the verification steps stay with the named drafting task.